Drax

An embeddable distributed k/v store for Go

Drax is a distributed k/v store based on raft (specifically https://github.com/hashicorp/raft) The intent is to be relatively light-weight and embeddable, while implementing the API from https://github.com/docker/libkv

Drax is NOT intended for production use. Use at your own risk. This should be considered ALPHA quality.

Boltdb is used to persist raft logs (and as such the K/V data)

Usage

For running the server, see cmd/kv.go as an example to get going.

Server Side

  // on node 1
  listener, _ := net.Listen("tcp", "10.0.0.1:2380")
  peerAddr := ""
  dialerFn := func(addr string, timeout time.Duration) (net.Conn, error) {
    return net.DialTimeout("tcp", addr, timeout)
  }
  cluster, _ := drax.New(listener, dialerFn, "/var/lib/drax", listener.Addr().String(), peerAddr)


  // on node 2
  listener, _ := net.Listen("tcp", "10.0.0.2:2380")
  peerAddr := "10.0.0.1:2380"
  dialerFn := func(addr string, timeout time.Duration) (net.Conn, error) {
    return net.DialTimeout("tcp", addr, timeout)
  }
  cluster, _ := drax.New(listener, dialerFn, "/var/lib/drax", listener.Addr().String(), peerAddr)

A node that does not specify a peer, and that does not already contain a peer in it's peer store will be setup in single node mode and made the leader. Once it is the leader it will exit single-node mode.

You can join a cluster by specifying any active peer's address, it does not need to be the leader.

Client side

  dialerFn := func(addr string, timeout time.Duration) (net.Conn, error) {
    return net.DialTimeout("tcp", addr, timeout)
  }
  retryTimeout := 10*time.Second
  client := client.New("10.0.0.1:2380", retryTimeout, dailerFn)
  kvPair, err := client.Get("/foo")

Requests to the K/V store can be sent to any peer and it will be forwarded to the leader. This implements libkv's Store interface.

You can also communicate with the k/v store directly from the cluster:

  kvPair, err := cluster.KVStore().Get("/foo")

This also implements libkv's Store interface.

###TODO:

• Add (more) tests
• Improve RPC semantics
• Look at using something other than JSON for encoding/decoding K/V messages, and RPC messages
• Implement cluster management, adding/removing nodes to the store cluster as needed